381457.fig.006a
(a)
381457.fig.006b
(b)
381457.fig.006c
(c)
Figure 6: Performance of each algorithm when trained on the overlapping squares task with (a) 𝐜 = [ 1 , 1 ] , (b) 𝑛 = 4 8 , and (c) 𝐩 = [ 0 . 1 , 0 . 1 ] . Results are shown for three different versions of each task; foreground bars show results when 𝐜 = [ 1 , 1 ] equals the number of image components, 𝑛 = 4 8 , and 𝐩 = [ 0 . 0 2 , 0 . 2 ] ; middle bars show results for 𝐜 = [ 0 . 1 , 1 ] , 𝚍 𝚒 𝚖 , and 𝚍 𝚒 𝚖 ; background bars show results for 𝑠 , 𝐩 , and 𝐜 , . Results are averaged over 10 trials for each condition. Plots in the left-hand column show the mean number of errors generated in the response of each network to 1000 test images. Each bar is subdivided into the proportion of false negatives (lighter, lower, section) and the proportion of false positives (darker, upper, section). Plots in the right-hand column show the mean number of components correctly represented by the synaptic weights learnt by each algorithm. Error bars show best and worst performance, across the 10 trials.